|
3D Multi-people Tracking
The 3D tracker uses wide baseline stereo to derive the 3D positions of
objects. At every frame, we measure the color distance between
all
possible pairings of tracks from the 2 views. We use the Bhattacharya
distance
between the normalized color histograms of the tracks. For each pair we
also measure the triangulation error, which is defined as the shortest
3D distance between the rays passing through the centroids of the
appearance
models in the two views. The triangulation error is generated using the
camera calibration data. To establish correspondence we minimize the
color
distance between the tracks from the view with the smaller number of
tracks
to the view with the larger number. This process can potentially lead
to
multiple tracks from one view being assigned to the same track in the
other.
We use the triangulation error to eliminate such multiple
assignments.
The triangulation error for the final correspondence is thresholded to
eliminate spurious matches that can occur when objects are just visible
in one of the two views. Once a correspondence is available at a given
frame, we now need to establish a match between the existing set of 3D
tracks and 3D objects present in the current frame. We use the
component
2D track identifiers of a 3D track and match them against the component
2D track identifiers of the current set of objects to establish the
correspondence.
The system also allows for partial matches, thus ensuring a continuous
3D track even when one of the 2D tracks fails. Thus the 3D tracker is
capable
of generating 3D position tracks of the centroid of each moving object
in the scene. It also has access to the 2D shape and color models from
the two views that make up the track. A demo of the 3D tracker is
shown.
All demo videos are in MPEG1 format. (video
5.8MB)
Other Research Areas:
|
Other
projects in